Overview

This page is the map. Read it once, top to bottom, and you will know what the package does, why it exists, and where every other page fits. Nothing here assumes you have used the package before.

What problem does this solve?

You have data on an outcome Y, a treatment or policy variable T, and a vector of covariates X. You believe the effect of T on Y is heterogeneous: it differs from person to person, depending on X. You want two things at once:

A flexible model of that heterogeneity. A neural network is good at this, because it can learn the shape of the effect β(X) without you guessing a functional form.
A valid confidence interval for a summary you care about, for example the average effect μ* = E[β(X)]. “Valid” means a 95% interval that actually contains the truth 95% of the time.

The catch is that you cannot have both for free. A neural network is regularized (early stopping, weight decay, dropout), and that regularization introduces bias into any summary you read off the fitted network. If you ignore the bias and just take the standard deviation of the network’s predictions, your confidence intervals are far too narrow. They cover the truth far less than 95% of the time.

The whole motivation in one table
Method	Coverage (target 95%)	SE ratio (target 1.0)
Naive neural network	8%	0.27
Influence-function corrected	95%	1.08

The “Naive” row is what you get if you treat the network’s output like a regression coefficient. Eight percent coverage means the interval you call “95%” is wrong nineteen times out of twenty. The corrected row is what this package delivers.

How does it fix the bias?

There are two procedures in the box. Both correct the regularization bias; they arrive at the correction differently.

        flowchart LR
    D["Data: Y, T, X"] --> NN["Neural net learns theta(X) = (alpha(X), beta(X))"]
    NN --> NAIVE["Naive summary: mean of beta(X), biased"]
    NN --> CORR["Add the influence-function correction term"]
    CORR --> VALID["Debiased estimate plus valid 95% CI"]

Influence-function procedure (Farrell, Liang, Misra). This is the default. It computes a per-observation correction term built from the model’s score and its Hessian, then averages the corrected quantity. The correction is exactly the amount of bias the regularization injected, removed. See Inference.
RieszNet procedure (Chernozhukov et al.). This learns the correction directly with a second neural-network head, without you deriving any Hessian by hand. It is the “automatic debiasing” route. See the RieszNet page.

Both rest on the same idea: machine learning and economic structure are complements, not substitutes. The network supplies the flexible heterogeneity; the structure (a known loss, a known target) supplies the object you can do honest inference on.

“The central idea is that machine learning methods and economic structure are complements, not substitutes. Machine learning methods alone predict well, but extrapolate nonsensically. Economic structure alone can produce robust inference, but may miss important heterogeneity that is visible in the data.” (Farrell, Liang, Misra, 2021)

What can you model?

The outcome can be almost any common type. Each one is a model (the package historically calls these “families”), and each has its own detailed page under Models.

Outcome type	Model
Continuous, unbounded	Linear, Gaussian
Binary (0/1)	Logit, Probit
Counts (0,1,2,…)	Poisson, Negative Binomial, Zero-Inflated Poisson
Positive, skewed	Gamma, Weibull
Censored	Tobit
Extreme values	Gumbel
Bounded in (0,1)	Beta
Discrete choice (3+ options)	Multinomial Logit
Quantiles	Quantile regression
Multiple treatments	Combinatorial
Before/after, treated/control	Difference-in-Differences, Panel Fixed-Effects DiD

What can you estimate?

The thing you want a confidence interval for is the target. The default target is the average effect E[β(X)], but the package ships many economic functionals, and you can supply your own (autodiff handles the derivatives).

Average marginal effect, price elasticity, willingness to pay, consumer welfare, dose-response, expected profit, tail probability, conditional variance, combinatorial treatment effects, and custom targets. Each is covered on its own page; targets are summarized in the API reference.

The three regimes (read this before you run anything serious)

The correction needs one nuisance object, the expected curvature of the loss, written Λ(x). How the package gets Λ(x) depends on your data, and that choice is called the regime.

Regime	When it applies	How `Λ(x)` is obtained
A	Randomized experiment with a known treatment distribution	Computed by Monte-Carlo integration (no estimation)
B	Linear / squared-error model	Closed form (the Hessian does not depend on the parameters)
C	Observational data, nonlinear model	Estimated with a separate regression, on a three-way data split

You usually do not pick the regime by hand. The package detects it from your inputs. The full explanation is in Estimation and Theory.

How this documentation is organized

Read the pages in menu order. They are written to be read linearly.

Section	What it gives you
Quick Start	Install, then a runnable example you can paste and run
Loading Data & Pre-Estimation	The exact shapes `Y`, `T`, `X` must have, and what to do before fitting
Models	One detailed page per outcome type
Estimation	How the network is fit and how `Λ(x)` is obtained (the machinery)
Inference	How standard errors and intervals are formed; the two procedures
Guide	Practical decisions: which model, which `Λ` method, how to read diagnostics
Theory	The formal walkthrough, in order, with the theorems
Replications	What we have validated against known truth, with the numbers
Simulation Studies	Head-to-head method comparisons (coverage, SE ratio, bias)
API Reference	Every function and result object
References	The papers, annotated

Who is this for?

Applied economists and data scientists who fit structural or causal models and need defensible standard errors, not just point predictions. You should be comfortable with a generalized linear model and with running Python. You do not need to know anything about influence functions going in; that is what the Theory section is for.